What are regularization images and how/when do I use them?

[tropic quiver]

title

[lost arrow]

Regularization images are images that don't show the training concept which are used to slow down or weaken the effect of training. It's usually adding outputs from the model to the training data to avoid overfitting.

Usually they aren't necessary if you already have good training parameters. I don't know of anyone who uses them actively.

[tropic quiver]

good answer, thanks! if anyone does use them regularly, mind giving some insight why you do?

[daring wagon]

They used to be useful for dreambooth back when that was the only way to train. Now they’re useless for the most part

[cursive schooner]

they aren't useless

[cursive schooner]

1. Prevents overfitting on style
2. Helps with overfitting of clothes
3. As a bonus, it becomes a cosplay LoRA
4. Gives better flexibility
5. Helps with multiple subjects LoRAs

[tropic quiver]

nice. what images do you use for regularization, and do you run them through a tagger?

[cursive schooner]

you can use either generated from the model or just from danbooru or something. And yes, you tag them

[tropic quiver]

does it have to be images from the lora or literally just whatever unrelated pics? also what's the ratio of regularization images you use per the amount of images used in the dataset?

[cursive schooner]

unrelated is fine, don't know about ratio

[tropic quiver]

ok out of 100 images in a dataset, how many regularization images do you use

or do you personally not count

[tropic quiver]

bump for visibility

[cursive schooner]

at least 100

[tropic quiver]

hum, okay